A Method for Blur and Similarity Transform Invariant Object Recognition

In this paper, we propose novel blur and similarity transform (i.e. rotation, scaling and translation) invariant features for the recognition of objects in images. The features are based on blur invariant forms of the log-polar sampled phase-only bispectrum and are invariant to centrally symmetric blur, including linear motion and out of focus blur. An additional advantage of using the phase-only bispectrum is the invariance to uniform illumination changes. According to our knowledge, the invariants of this paper are the first blur and similarity transform invariants in the Fourier domain. We have compared our features to the blur invariants based on complex image moments using simulated and real data. The moment invariants have not been evaluated earlier in the case of similarity transform. The results show that our invariants can recognize objects better in the presence of noise.